Detecting geological structures in seismic volumes using deep convolutional neural networks

Identifying the geological structures in seismic volumes is of great importance for oil and gas exploration. However, seismic data interpretation is a time consuming manual task even for experienced experts. In this thesis we propose an automatic method based on 2D and 3D convolutional neural networks (CNN) to detect geophysical structures such as channels and faults in 3D seismic volumes. We apply CNN as a local classifier to 2D and 3D patches around every voxel in the seismic volume in order to perform semantic segmentation. The models trained on patches from orthogonal views are then ensembled to improve the classification accuracy. We have conducted extensive experiments on the Parihaka and F3 volumes and presented detailed results. With the help of regularization techniques, our models generalize well to new data, despite the fact that only a small training set generated from weakly labeled data is provided. Our experiments show that convolutional neural networks trained on raw pixel intensities are capable of achieving high-quality segmentation results in the seismic interpretation field that requires specific domain knowledge.